huggingface--transformers/docs/source/en/model_doc/voxtral_realtime.md

This model was contributed to Hugging Face Transformers on 2026-02-16.

VoxtralRealtime

VoxtralRealtime is a streaming speech-to-text model from Mistral AI, designed for real-time automatic speech recognition (ASR). Unlike the offline Voxtral model which processes complete audio files, VoxtralRealtime is architected for low-latency, incremental transcription by processing audio in chunks as they arrive.

The model combines an audio encoder with a Mistral-based language model decoder, using time conditioning embeddings and causal convolutions with padding caches to enable efficient streaming inference.

Usage

Offline Transcription

For transcribing complete audio files, use the processor and model directly. The generation length is automatically determined from the audio length.

from datasets import load_dataset

from transformers import AutoProcessor, VoxtralRealtimeForConditionalGeneration


repo_id = "mistralai/Voxtral-Mini-4B-Realtime-2602"

processor = AutoProcessor.from_pretrained(repo_id)
model = VoxtralRealtimeForConditionalGeneration.from_pretrained(repo_id, device_map="auto")

ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
audio = ds[0]["audio"]["array"]

inputs = processor(audio, return_tensors="pt").to(model.device)
inputs = inputs.to(model.device, dtype=model.dtype)

outputs = model.generate(**inputs)
decoded_outputs = processor.batch_decode(outputs, skip_special_tokens=True)

print(decoded_outputs[0])

Batched Offline Transcription

Multiple audio samples can be transcribed in a single forward pass:

from datasets import load_dataset

from transformers import AutoProcessor, VoxtralRealtimeForConditionalGeneration


repo_id = "mistralai/Voxtral-Mini-4B-Realtime-2602"

processor = AutoProcessor.from_pretrained(repo_id)
model = VoxtralRealtimeForConditionalGeneration.from_pretrained(repo_id, device_map="auto")

ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
audio = [ds[i]["audio"]["array"] for i in range(2)]

inputs = processor(audio, return_tensors="pt").to(model.device)
inputs = inputs.to(model.device, dtype=model.dtype)

outputs = model.generate(**inputs)
decoded_outputs = processor.batch_decode(outputs, skip_special_tokens=True)

for decoded_output in decoded_outputs:
    print(decoded_output)

Streaming Transcription

Note

This is an experimental feature and the API is subject to change.

For real-time transcription, audio is split into chunks following:

from threading import Thread

import numpy as np
from datasets import load_dataset

from transformers import (
    TextIteratorStreamer,
    VoxtralRealtimeForConditionalGeneration,
    VoxtralRealtimeProcessor,
)


model_id = "mistralai/Voxtral-Mini-4B-Realtime-2602"
processor = VoxtralRealtimeProcessor.from_pretrained(model_id)
model = VoxtralRealtimeForConditionalGeneration.from_pretrained(model_id, device_map="cuda:0")

ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
audio = ds[0]["audio"]["array"]
# Manually pad the audio to account for right padding tokens required by the model
xaudio = np.pad(audio, (0, processor.num_right_pad_tokens * processor.raw_audio_length_per_tok))

first_chunk_inputs = processor(
    audio[:processor.num_samples_first_audio_chunk],
    is_streaming=True,
    is_first_audio_chunk=True,
    return_tensors="pt"
)
first_chunk_inputs.to(model.device, dtype=model.dtype)

def input_features_generator():
    yield first_chunk_inputs.input_features

    mel_frame_idx = processor.num_mel_frames_first_audio_chunk
    hop_length = processor.feature_extractor.hop_length
    win_length = processor.feature_extractor.win_length

    start_idx = mel_frame_idx * hop_length - win_length // 2
    end_idx = start_idx + processor.num_samples_per_audio_chunk

    while (end_idx:=start_idx + processor.num_samples_per_audio_chunk) < audio.shape[0]:
        inputs = processor(
            audio[start_idx:end_idx],
            is_streaming=True,
            is_first_audio_chunk=False,
            return_tensors="pt"
        )
        inputs.to(model.device, dtype=model.dtype)
        yield inputs.input_features

        mel_frame_idx += processor.audio_length_per_tok
        start_idx = mel_frame_idx * hop_length - win_length // 2

streamer = TextIteratorStreamer(processor.tokenizer, skip_special_tokens=True, clean_up_tokenization_spaces=True)
generate_kwargs = {
    "input_ids": first_chunk_inputs.input_ids,
    "input_features": input_features_generator(),
    "num_delay_tokens": first_chunk_inputs.num_delay_tokens,
    "streamer": streamer,
}
thread = Thread(target=model.generate, kwargs=generate_kwargs)
thread.start()

# Iterate over the streamer to get text chunks as they are generated
print("Model output (streaming):", end=" ", flush=True)
for text_chunk in streamer:
    print(text_chunk, end="", flush=True)

This model was contributed by Eustache Le Bihan.

VoxtralRealtimeConfig

autodoc VoxtralRealtimeConfig

VoxtralRealtimeEncoderConfig

autodoc VoxtralRealtimeEncoderConfig

VoxtralRealtimeTextConfig

autodoc VoxtralRealtimeTextConfig

VoxtralRealtimeFeatureExtractor

autodoc VoxtralRealtimeFeatureExtractor

VoxtralRealtimeProcessor

autodoc VoxtralRealtimeProcessor - call

VoxtralRealtimeEncoder

autodoc VoxtralRealtimeEncoder - forward

VoxtralRealtimeModel

autodoc VoxtralRealtimeModel - forward

VoxtralRealtimeForConditionalGeneration

autodoc VoxtralRealtimeForConditionalGeneration - forward - get_audio_features